‘Electronic tongue’ identifies brands of beer with 81.9% accuracy
February 4, 2014
Spanish researchers have managed to distinguish between different varieties of beer using an “electronic tongue,” with an accuracy of 81.9%.
Scientists at the Autonomous University of Barcelona used an array of 21 sensors formed from ion-selective electrodes, including some with response to cations (ammonium, sodium), others with response to anions (nitrate, chloride, etc.), and electrodes with generic (unspecified) responses.
The authors recorded the multidimensional response generated by the sensors and applied supervised learning and linear discriminant analysis.
The researchers said these tools could one day give robots a sense of taste, and even supplant panels of tasters in the food industry to improve the quality and reliability of products for consumption.
Will wine tasting be next?
Abstract of Food Chemistry paper
In this work, an electronic tongue (ET) system based on an array of potentiometric ion-selective electrodes (ISEs) for the discrimination of different commercial beer types is presented. The array was formed by 21 ISEs combining both cationic and anionic sensors with others with generic response. For this purpose beer samples were analyzed with the ET without any pretreatment rather than the smooth agitation of the samples with a magnetic stirrer in order to reduce the foaming of samples, which could interfere into the measurements. Then, the obtained responses were evaluated using two different pattern recognition methods, principal component analysis (PCA), which allowed identifying some initial patterns, and linear discriminant analysis (LDA) in order to achieve the correct recognition of sample varieties (81.9% accuracy). In the case of LDA, a stepwise inclusion method for variable selection based on Mahalanobis distance criteria was used to select the most discriminating variables. In this respect, the results showed that the use of supervised pattern recognition methods such as LDA is a good alternative for the resolution of complex identification situations. In addition, in order to show an ET quantitative application, beer alcohol content was predicted from the array data employing an artificial neural network model (root mean square error for testing subset was 0.131 abv).